LIVE
Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|
Research

A Log(N)-Questions Game Tests Communication Efficiency of Frontier Models

Six frontier models were pitted against each other in a '20 questions' style game requiring them to identify a target among Wikipedia excerpts using a minimal number of yes/no questions. The study reveals clear performance gaps and a surprisingly stable information efficiency.

September 16, 20263 min readPublished byarXiv

A research team tested six frontier language models — Claude Opus 5, GLM-5.3, GPT-5.6 Sol, Grok 4.6, Gemini 3.8 Flash, and Kimi K3 — in a two-agent game where a questioner must identify, among N Wikipedia lead paragraphs, a target known only to an answerer, using exactly log2(N) yes/no questions. Because both roles are played by the same model, the setup measures how well a system communicates with itself under information asymmetry. The experiment covered 408 games across document sets ranging from 4 to 1024 paragraphs, at a total API cost of $363.

The results show a clear gap for Claude Opus 5, which wins only 28 of its 68 games, compared with 45 to 56 wins for the other five models, whose performance remains closely bunched. Pooling these five models, win rate declines with set size in a near-perfect correlation (r = -0.973), fitting a simple model where success rate equals p raised to the power of log2(N), with p estimated at 0.928 per round.

Failures split roughly evenly between answer errors and discrimination failures at the final guess. A detailed review of the weakest model's unanimous mistakes found that 32 of 34 involved incorrect "No" answers about properties stated in the document's opening sentence, despite explicit instructions warning against defaulting to "No." By contrast, models almost never name a document that their own prior answers had already ruled out.

Information efficiency, estimated from answer balance, correlates strongly with win rate (r = +0.88). Only two models manage to extract a full bit of information per question, by adopting a partitioning strategy based on document titles — an approach absent below 32 candidates but used in roughly a quarter of questions above that threshold. Reasoning-token usage varies by a factor of 4.5 across models, with little relationship to success: reasoning traces grow longer as the candidate pool shrinks, without a matching improvement in reliability. The findings suggest that extended reasoning alone does not translate into efficient communication under information asymmetry, a relevant concern for the design of multi-agent AI systems.

Tags
benchmarkmulti-agentllm-evaluationinformation-theoryself-playreasoning

Read also